caiotbr/latent-raw-agent-samples

Latent Raw Agent: sample data

Sample inputs and precomputed pass outputs for Atikum-AI/latent-raw-agent, an orchestrator agent that turns 8-bit AI video into gradeable 32-bit ACES masters. The agent's SAMPLE_MODE replays these files, so judges and users can run it with no GPU and no API keys.

All four clips are AI-generated by Kyo Silva (Atikum). One key frame per clip.

clip_id	key frame	size	items (SAM 3.1 masks)
clip01_car_street_sun	48 (24 fps)	2200×944	car, sky, buildings, road
clip02_lantern_table	48 (24 fps)	1280×720	lantern_flame, girl, table, hands
clip03a_canoe_dawn	48 (24 fps)	1280×720	person_young, person_old, water, sky
clip03b_jungle_warm	60 (30 fps)	1920×1080	person_young, person_old, forest

Files per clip (<clip_id>/)

File	What it is	How it was made
clip.json	key frame, colour space (AP1 = ACEScg), source file
source.png	the key frame, 8-bit sRGB, as the video model made it	ffmpeg frame export
masks/<item>.png	one matte per item, frame-sized, white = inside	SAM 3.1 (sam3.1_multiplex_fp16) in ComfyUI, one text prompt per item (tools/sam_masks.py)
synth.png	pixels the HDR expansion had to synthesize: 8-bit source clipped (any channel ≥ 250), dilated 2 px	tools/finish.py
expand_hdr.exr	key frame after SDR→HDR expansion, ACEScg linear, half float	Lightricks LTX-2.5 SDR-to-HDR IC-LoRA on a 17-frame segment around the key frame (1280 wide), resized back to the source size (tools/hdr_graphs.json)
balance.exr	key frame after the balance pass, ACEScg linear	tools/finish.py: plain linear gains, see balance.json
balance.json	every gain applied and the numbers it was computed from
evidence.json	the agent's measured rows (per-item clipping, luminance, colour balance, range)	agent/samples.py in the agent repo

The balance pass

• Single clips: one exposure gain so the subject's median luminance reaches its target (car paint 0.09, skin 0.18), capped at ±1.5 stops. Items protected by a colourist ruling keep their values: in clip02 the lantern flame (practical_flames_hot: practical flames may clip, never pull them down).
• The pair clip03a → clip03b: per-channel gains on clip03b: the geometric mean over the two people of (mean RGB in clip03a / mean RGB in clip03b), the statistic the agent checks. Result: the people match within 0.005 stops.

Expanded range is synthesized by the model, never "recovered" from the 8-bit source.

Licence

Provided for evaluating and reproducing the Latent Raw Agent. All rights reserved by Atikum otherwise; see LICENSE.